Qwen Image orbit LoRA beats base model on view synthesis: 0.794 vs 0.731 alpha IoU
linoy_tsaban · x · 2026-09-28
A Qwen Image orbit LoRA trained on 2K synthetic renders from Google Scanned Objects supports 23 relative camera instructions (45°/90°/135° rotations plus elevation), letting the model generate unseen sides on command.
- Beats the base model: 0.794 vs 0.731 alpha IoU, with the biggest gains on 90° rotations where the base model tends to hallucinate hidden sides
- RGBA in/out: transparent images stay transparent, no matting needed
- Live demo available on Hugging Face Spaces
More from Multimodal
- Agentic filmmaking experiment: GPT-6 Astra edits, Grok Imagine renders, Suno v6 scores — TinfoilTricorn · 2026-09-29
- Fizgig 6.6 adds edit-LoRA training for Qwen Image 2.1, learns grading from 40 pairs — shootthesound · 2026-09-29
- Deforum keeps going and going: an endless morphing chair animation — makeitrad1 · 2026-09-29
- GPT Image 2.5 vs open-source Krea 2 Turbo: same-prompt test highlights OpenAI's heavy censorship — Due_Research9042 · 2026-09-29
- Hugging Face release: Yue2-based quiet-storm model for 90s Jazz-Soul vibes — -becausereasons- · 2026-09-29
- One prompt, six shots: multi-shot video generation shows cinematic chops — minchoi · 2026-09-29